Abstract geometric composition illustrating everyone became an ai portrait and somebody else paid for it

December 2022 ยท AI & Technology

Everyone Became an AI Portrait and Somebody Else Paid For It

The app was fun, the outputs were flattering, and the training data belonged to working illustrators who were not asked.

Four of my group chats turned into a portrait gallery this week. Everyone I know is suddenly an elf, an astronaut, a Renaissance oil painting with cheekbones they did not previously have. You pay a few dollars, upload twenty selfies, wait twenty minutes, and Lensa hands you back fifty flattering versions of yourself.

I understand the appeal completely. Nobody has liked a photo of themselves since roughly 2019, and here is a machine that makes you look like the protagonist.

Then a friend of mine who illustrates book covers for a living posted her set. Not avatars. Screenshots of outputs from these models with the smeared remains of a signature still sitting in the bottom corner.

That detail has been circulating among illustrators for months now, and it stopped being funny some time ago.

Every one of those portraits is a small withdrawal from an account that working illustrators funded and were never asked to open.

The mechanics are not secret. Lensa is built on Stable Diffusion, which was trained on a scrape of billions of captioned images pulled off the open web, including the portfolios, the ArtStation pages, the client work, the personal projects. Living artists have found their names functioning as prompt modifiers, a style available on request, and none of them signed anything or received anything. The output is not a copy of any single piece. It is something more slippery: a statistical impression of thousands of careers, sold back at four dollars a set.

Living artists have found their names functioning as prompt modifiers, a style available on request, and none of them signed anything or received anything.

Here is the part that matters for anyone building a brand, because I have already seen three decks this quarter proposing generated imagery for Q1 campaigns.

One, provenance is now a brand attribute whether you want it to be or not. Your audience contains illustrators, photographers, and people who follow illustrators. They can tell. Two, running it through a vendor does not launder it. If your agency delivers assets from a model with unlicensed training data, that is your logo underneath, not theirs. Three, the actual cost of commissioning a human is smaller than the cost of explaining yourself afterward, and the human gives you something the model structurally cannot, which is a point of view nobody else can generate.

The tools are extraordinary and they are not going away. Neither is the question of who paid for them.

There is a print above my desk that cost me two hundred dollars and took the artist three weeks, a rainy Hawthorne street corner with the light doing something impossible. I know her name. I know what she charges. Truffles sleeps directly beneath it every afternoon, unmoved by the ethics, warmed by the radiator, which is roughly the level of engagement I expect from her on any topic.